v10.3.0+: backfill fails with MetadataValue conversion error via chroma-mcp

[GigiTiti-Kai]

Describe the bug

After upgrading to v10.3.0+ (which switched from JS Chroma bindings to Python-based chroma-mcp via uvx), backfill of existing observations to ChromaDB fails for projects with large datasets. The error occurs in chroma_add_documents when metadata contains values that ChromaDB cannot accept as MetadataValue (only str/int/float/bool are supported).

New individual observation syncs work fine — the issue is specific to the bulk backfill code path.

Error message:

[CHROMA_SYNC] Backfilling observations {project=project-A, missing=1760, existing=0, total=1760}
[CHROMA_SYNC] Backfill failed {project=project-A} chroma-mcp tool "chroma_add_documents" returned error:
  Error executing tool chroma_add_documents: Failed to add documents to collection 'cm__claude-mem':
  argument 'metadatas': Cannot convert Python object to MetadataValue

Affected projects (same error):

• project-A (1760 observations) — failed
• project-B — failed
• project-C — failed
• project-D — failed (different error: JSON Parse error: Unexpected identifier "server")
• Small projects (project-E, project-F, project-G, etc.) — succeeded

Root cause:

addDocuments() sends batches of documents with metadata to chroma_add_documents. ChromaDB metadata only supports scalar types (str, int, float, bool). If any observation in a batch contains metadata with arrays, nested objects, or null values, the entire batch fails.

The previous JS-based Chroma integration likely handled type coercion internally. The new chroma-mcp bridge passes metadata as-is to Python chromadb, which strictly validates types.

Small projects succeed because they don't happen to contain observations with non-scalar metadata. Large projects inevitably have some.

Steps to reproduce

1. Have existing observation data from pre-v10.3.0
2. Upgrade to v10.3.0+ (or delete ~/.claude-mem/chroma/ to force re-backfill)
3. Wait for backfill to trigger on session start
4. Check logs: large projects fail with Cannot convert Python object to MetadataValue

Suggested fix

Sanitize metadata before passing to chroma_add_documents — convert non-scalar values to strings:

// In addDocuments() or the metadata construction step:
const sanitizeMetadata = (meta) => {
  const result = {};
  for (const [key, value] of Object.entries(meta)) {
    if (value === null || value === undefined) continue;
    if (typeof value === 'object') {
      result[key] = JSON.stringify(value);
    } else {
      result[key] = value;
    }
  }
  return result;
};

Environment

• claude-mem: 10.3.1
• chroma-mcp: 0.2.6 (via uvx)
• Platform: WSL2 (Ubuntu 24.04, kernel 6.6.87.2)
• Total observations across all projects: ~12,000+

Additional context

• Individual observation sync (non-backfill) works correctly — likely because new observations are constructed with scalar-only metadata
• project-D fails with a different error (JSON Parse error), which may be a separate data corruption issue
• Related: v10.3.0+: chroma-mcp SSL default breaks CHROMA_MODE=remote with local HTTP server #1182 (SSL default issue in remote mode), v10.2.5: cache directory missing node_modules — semantic search broken on marketplace install #1166 (node_modules missing in cache)

[elador13]

Experiencing the same issue on Windows 11 (native, not WSL).

Environment:

• claude-mem: 10.3.1
• chroma-mcp: via uvx (Python 3.13)
• Platform: Windows 11

Additional context:
On Windows, I also had to create a uvx.cmd shim because claude-mem hardcodes uvx.cmd for Windows but the uv standalone installer only provides uvx.exe. Once that was resolved, the backfill ran but hit the same MetadataValue error:

[CHROMA_SYNC] Backfilling observations {project=***, missing=221, existing=1, total=222}
[CHROMA_SYNC] Backfilling summaries {project=***, missing=27, existing=0, total=27}
[CHROMA_SYNC] Backfilling user prompts {project=***, missing=709, existing=0, total=709}
[CHROMA_SYNC] Backfill failed {project=***} chroma-mcp tool "chroma_add_documents" returned error: Error executing tool chroma_add_documents: Failed to add documents to collection 'cm__claude-mem': argument 'metadatas': Cannot convert Python object to MetadataValue

Observations backfilled successfully, but summaries and user prompts failed. The suggested metadata sanitization fix looks correct — would love to see this in the next patch.